You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
keicoon15 edited this page Aug 21, 2018
·
1 revision
우선 dataset으로 celebA 이미지를 가공했습니다.
64x64 size에 맞게 이미지를 resize하고 얼굴 부분을 얻기 위해 croping이 작업된 이미지를 pre-processing하고 example.mnist.dataset의 인터페이스를 가져와 shuffle_batch 함수를 만들었습니다.
VAE 모델에도 수정이 이뤄졌는데 기존에 dense로 구성했던 encoder, decoder 부분을 conv2d layer로 대체했습니다. 이전 VAE 모델에서 사용한 mnist 데이터에 비해 celebA는 (r, g, b) 3 체널로 구성되어 있고 value 값도 variation이 다양하기 때문에 복잡한 모델을 필요로 했습니다. 그리고 타겟 이미지가 가지는 class 수도 증가했기 때문에 z(latent-attention)도 값을 늘려 실험을 진행했습니다. 마지막으로 loss 계산에서 recon_loss 계산에 MSE를 적용했습니다.